2  Preparing Raster Data

3 Load data on study area

3.1 Soil samples

Code
# Load soil data from sampling locations
bern_data <- readr::read_csv(
  here::here("data-raw/soildata/berne_soil_sampling_locations.csv")
  )

# Display data
head(bern_data) |> knitr::kable()
site_id_unique timeset x y dataset dclass waterlog.30 waterlog.50 waterlog.100 ph.0.10 ph.10.30 ph.30.50 ph.50.100
4_26-In-005 d1968_1974_ptf 2571994 1203001 validation poor 0 0 1 6.071733 6.227780 7.109235 7.214589
4_26-In-006 d1974_1978 2572149 1202965 calibration poor 0 1 1 6.900000 6.947128 7.203502 7.700000
4_26-In-012 d1974_1978 2572937 1203693 calibration moderate 0 1 1 6.200000 6.147128 5.603502 5.904355
4_26-In-014 d1974_1978 2573374 1203710 validation well 0 0 0 6.600000 6.754607 7.200000 7.151129
4_26-In-015 d1968_1974_ptf 2573553 1203935 validation moderate 0 0 1 6.272715 6.272715 6.718392 7.269008
4_26-In-016 d1968_1974_ptf 2573310 1204328 calibration poor 0 0 1 6.272715 6.160700 5.559031 5.161655

The dataset on soil samples from Bern holds 13 variables for 1052 entries (more information here):

  • site_id_unique: The location’s unique site id.

  • timeset: The sampling year and information on sampling type for soil pH (no label: CaCl\(_2\) laboratory measurement, field: indicator solution used in field, ptf: H\(_2\)O laboratory measurement transferred by pedotransfer function).

  • x: The x (easting) coordinates in meters following the (CH1903/LV03) system.

  • y: The y (northing) coordinates in meters following the (CH1903/LV03) system.

  • dataset: Specification whether a sample is used for model calibration or validation (this is based on randomization to ensure even spatial coverage).

  • dclass: Soil drainage class

  • waterlog.30, waterlog.50, waterlog.100: Specification whether soil was water logged at 30, 50, or 100 cm depth (0 = No, 1 = Yes).

  • ph.0.10, ph.10.30, ph.30.50, ph.50.100: Average soil pH between 0-10, 10-30, 30-50, and 50-100 cm depth.

3.1.1 Covariate data

Now, let’s load all the covariates that we want to produce our soil maps with.

Code
# Get a list with the path to all raster files
list_raster <- 
  base::list.files(
    here::here("data-raw/geodata/covariates/"),
    full.names = T
  )

# Take a random subset
set.seed(123)
list_raster_subset <- list_raster |> sample(15)

# Display data (lapply to clean names)
lapply(
  list_raster_subset, 
  function(x) sub(".*/(.*)", "\\1", x)
  ) |> 
  unlist() |> 
  head(5) |> 
  print()
[1] "Se_curvplan2m_fmean_50c.tif" "Se_tpi_2m_5c.tif"           
[3] "Se_e_aspect2m_5c.tif"        "mt_tt_y.tif"                
[5] "Se_SCA2m.tif"               

The output above shows the first five raster files with rather cryptic names. For this tutorial, we do not need to understand all of the 91 rasters files that we are loading here but in a scientific context, you should know your data better of course. Let’s look at one of these raster files to get a better feeling for our data. Specifically, let’s look at the slope profile at 2m resolution:

Code
# Load a raster file as example: Picking the slope profile at 2m resolution
raster_example <- terra::rast(list_raster[74])
raster_example
class       : SpatRaster 
dimensions  : 986, 2428, 1  (nrow, ncol, nlyr)
resolution  : 20, 20  (x, y)
extent      : 2568140, 2616700, 1200740, 1220460  (xmin, xmax, ymin, ymax)
coord. ref. : CH1903+ / LV95 
source      : Se_slope2m.tif 
name        : Se_slope2m 
min value   :    0.00000 
max value   :   85.11286 

As shown in the output, a raster object has the following properties (among others, see ?terra::rast):

  • class: The class of the file, here a SpatRaster.

  • dimensions: The number of rows, columns, years (if temporal encoding).

  • resolution: The resolution of the coordinate system, here it is 20 in both axes.

  • extent: The extent of the coordinate system defined by min and max values on the x and y axes.

  • coord. ref.: Reference coordinate system. Here, the raster is encoded using the LV95 geodetic reference system from which the projected coordinate system CH1903+ is derived.

  • source: The name of the source file.

  • names: The name of the raster file (mostly the file name without file-specific ending)

  • min value: The lowest value of all cells.

  • max value: The highest value of all cells.

Tip

The code chunks filtered for a random sub-sample of 15 variables. As described in Chapter 5, your task will be to investigate all covariates and find the ones that can best be used for your modelling task.

4 Mapping the study area

Now, let’s look at a visualisation of this raster file. Since we selected the slope at 2m resolution, we expect a relief-like map with a color gradient that indicates the steepness of the terrain.

Code
# Plot raster example
terra::plot(raster_example)

Code
# To have some more flexibility, we can plot this in the ggplot-style as such:
ggplot2::ggplot() +
  tidyterra::geom_spatraster(data = raster_example) +
  ggplot2::scale_fill_viridis_c(
    na.value = NA,
    option = "magma",
    name = "Slope (%) \n"
    ) +
  ggplot2::theme_bw() +
  ggplot2::scale_x_continuous(expand = c(0, 0)) +  # avoid gap between plotting area and axis
  ggplot2::scale_y_continuous(expand = c(0, 0)) +
  ggplot2::labs(title = "Slope of the Study Area")

Tip

Note that the second plot has different coordinates than the upper one. That is because the data was automatically projected to the World Geodetic System (WGS84, ESPG: 4326).

This looks already interesting but we can put our data into a bit more context. For example, a larger map background would be useful to get a better orientation of our location. Also, it would be nice to see where our sampling locations are and to differentiate these locations by whether they are part of the calibration or validation dataset. Bringing this all together requires some more understanding of plotting maps in R. So, don’t worry if you do not understand everything in the code chunk below and enjoy the visualizations:

Code
# To get our map working correctly, we have to ensure that all the input data
# is in the same coordinate system. Since our Bern data is in the Swiss 
# coordinate system, we have to transform the sampling locations to the 
# World Geodetic System first.
# To look up EPSG Codes: https://epsg.io/
# World Geodetic System 1984:  4326
# Swiss CH1903+ / LV95: 2056

# For the raster:
r <- terra::project(raster_example, "+init=EPSG:4326")

# Let's make a function for transforming the sampling locations:
change_coords <- function(data, from_CRS, to_CRS) {
  
  # Check if data input is correct
  if (!all(names(data) %in% c("id", "lat", "lon"))) {
    stop("Input data needs variables: id, lat, lon")
  }
  
  # Create simple feature for old CRS
  sf_old_crs <- sf::st_as_sf(data, coords = c("lon", "lat"), crs = from_CRS)
  
  # Transform to new CRS
  sf_new_crs     <- sf::st_transform(sf_old_crs, crs = to_CRS)
  sf_new_crs$lat <- sf::st_coordinates(sf_new_crs)[, "Y"]
  sf_new_crs$lon <- sf::st_coordinates(sf_new_crs)[, "X"]
  
  sf_new_crs <- sf_new_crs |> dplyr::as_tibble() |> dplyr::select(id, lat, lon)
  
  # Return new CRS
  return(sf_new_crs)
}

# Transform dataframes
coord_cal <- 
  bern_data |> 
  dplyr::filter(dataset == "calibration") |> 
  dplyr::select(site_id_unique, x, y) |> 
  dplyr::rename(id = site_id_unique, lon = x, lat = y) |> 
  change_coords(
    from_CRS = 2056, 
    to_CRS = 4326
    )

coord_val <- 
  bern_data |> 
  dplyr::filter(dataset == "validation") |> 
  dplyr::select(site_id_unique, x, y) |> 
  dplyr::rename(id = site_id_unique, lon = x, lat = y) |> 
  change_coords(
    from_CRS = 2056, 
    to_CRS = 4326
    )
Code
# Notes: 
# - This code may only work when installing the development branch of {leaflet}:
# remotes::install_github('rstudio/leaflet')
# - You might have to do library(terra) for R to find functions needed in the backend
library(terra)

# Let's get a nice color palette now for easy reference
pal <- leaflet::colorNumeric(
  "magma",
  terra::values(r),
  na.color = "transparent"
  )

# Next, we build a leaflet map
leaflet::leaflet() |> 
  # As base maps, use two provided by ESRI
  leaflet::addProviderTiles(leaflet::providers$Esri.WorldImagery, group = "World Imagery") |>
  leaflet::addProviderTiles(leaflet::providers$Esri.WorldTopoMap, group = "World Topo") |>
  # Add our raster file
  leaflet::addRasterImage(
    r,
    colors = pal,
    opacity = 0.6,
    group = "raster"
    ) |>
  # Add markers for sampling locations
  leaflet::addCircleMarkers(
    data = coord_cal,
    lng = ~lon,  # Column name for x coordinates
    lat = ~lat,  # Column name for y coordinates
    group = "training",
    color = "black"
  ) |>
    leaflet::addCircleMarkers(
    data = coord_val,
    lng = ~lon,  # Column name for x coordinates
    lat = ~lat,  # Column name for y coordinates
    group = "validation",
    color = "red"
  ) |>
  # Add some layout and legend
  leaflet::addLayersControl(
    baseGroups = c("World Imagery","World Topo"),
    position = "topleft",
    options = leaflet::layersControlOptions(collapsed = FALSE),
    overlayGroups = c("raster", "training", "validation")
    ) |>
  leaflet::addLegend(
    pal = pal,
    values = terra::values(r),
    title = "Slope (%)")
Note

This plotting example is based to the one shown in the AGDS 2 tutorial “Handful of Pixels” on phenology. More information on using spatial data in R can be found there in the Chapter on Geospatial data in R.

That looks great! At first glance, it is a bit crowded but once you scroll in you can investigate our study area quite nicely. You can check whether the slope raster file makes sense by comparing it against the base maps. Can you see how cliffs along the Aare river, hills, and even gravel quarries show high slopes? We also see that our validation dataset is nicely distributed across the area covered by the training dataset.

Now that we have played with a few visualizations, let’s get back to preparing our data. The {terra} package comes with the very useful tool to stack multiple raster on top of each other, if they are of the same spatial format. To do so, we just have to feed in the vector of file names list_raster_subset:

Code
# Load all files as one batch
all_rasters <- terra::rast(list_raster_subset)
all_rasters
class       : SpatRaster 
dimensions  : 986, 2428, 15  (nrow, ncol, nlyr)
resolution  : 20, 20  (x, y)
extent      : 2568140, 2616700, 1200740, 1220460  (xmin, xmax, ymin, ymax)
coord. ref. : CH1903+ / LV95 
sources     : Se_curvplan2m_fmean_50c.tif  
              Se_tpi_2m_5c.tif  
              Se_e_aspect2m_5c.tif  
              ... and 12 more source(s)
names       : Se_cu~n_50c, Se_tpi_2m_5c, Se_e_~2m_5c, mt_tt_y, Se_SCA2m, Se_cu~m_s60, ... 
min values  :  -0.8545234,    -19.00460,          -1,    35.0,        2,   -2.087684, ... 
max values  :   1.4742721,     25.30661,           1,   109.8, 21962074,    2.255710, ... 

Now, we do not want to have the covariates’ data from all cells in the raster file. Rather, we want to reduce our stacked rasters to the x and y coordinates for which we have soil sampling data. We can do this using the terra::extract() function. Then, we want to merge the two dataframes of soil data and covariates data by their coordinates. Since the order of the covariate data is the same as the Bern data, we can simply bind their columns with cbind():

Code
# Extract coordinates from sampling locations
sampling_xy <- bern_data |> dplyr::select(x, y)

# From all rasters, extract values for sampling coordinates
covar_data <- 
  terra::extract(all_rasters,  # The raster we want to extract from
                 sampling_xy,  # A matrix of x and y values to extract for
                 ID = FALSE    # To not add a default ID column to the output
                 )

final_data <- cbind(bern_data, covar_data)
head(final_data) |> knitr::kable() 
site_id_unique timeset x y dataset dclass waterlog.30 waterlog.50 waterlog.100 ph.0.10 ph.10.30 ph.30.50 ph.50.100 Se_curvplan2m_fmean_50c Se_tpi_2m_5c Se_e_aspect2m_5c mt_tt_y Se_SCA2m Se_curvprof2m_s60 Se_e_aspect25m Se_curvprof2m_s7 tsc25_18 Se_curv2m_std_50c Se_slope2m_fmean_5c Se_n_aspect2m_5c mrrtf25 Se_slope2m_std_50c Se_curv2m_std_5c
4_26-In-005 d1968_1974_ptf 2571994 1203001 validation poor 0 0 1 6.071733 6.227780 7.109235 7.214589 -0.0445323 -0.0583917 -0.7929600 98 16.248077 -0.0506380 -0.9702092 -0.0732220 0.3332805 2.9204133 0.6683306 -0.6084610 0.0184651 0.8815832 1.1769447
4_26-In-006 d1974_1978 2572149 1202965 calibration poor 0 1 1 6.900000 6.947128 7.203502 7.700000 -0.0501855 0.0180000 0.8753148 98 3.357315 -0.1005311 0.5683194 -0.4981292 0.3395441 3.8783867 0.9857153 0.4801802 0.0544361 1.0152543 4.3162045
4_26-In-012 d1974_1978 2572937 1203693 calibration moderate 0 1 1 6.200000 6.147128 5.603502 5.904355 -0.0079620 -0.0145804 -0.3866692 98 11.330072 -0.0125471 -0.6987815 -0.0052359 0.4455501 0.7022317 0.5300468 -0.9221049 3.6830916 0.4975456 0.4170935
4_26-In-014 d1974_1978 2573374 1203710 validation well 0 0 0 6.600000 6.754607 7.200000 7.151129 -0.0301961 -0.0085602 -0.8657616 98 42.167496 -0.0400928 -0.8485889 0.0529446 0.4483251 1.5150748 0.8635756 -0.4998477 0.0075817 0.5767300 0.2413423
4_26-In-015 d1968_1974_ptf 2573553 1203935 validation moderate 0 0 1 6.272715 6.272715 6.718392 7.269008 -0.0179657 0.0061576 -0.8864348 98 5.479310 -0.0308456 -0.8918364 0.0929077 0.3974232 3.6032522 1.2098556 0.4614536 0.0007469 2.7759163 1.8169731
4_26-In-016 d1968_1974_ptf 2573310 1204328 calibration poor 0 0 1 6.272715 6.160700 5.559031 5.161655 -0.0049875 0.0034276 0.5905659 98 13.499996 -0.0052602 -0.8766075 0.0867119 0.4278295 1.5897882 0.8515157 0.8054439 0.0128017 1.2163279 0.8171870

Great that worked without problems!

Now, not all our covariates may be continuous variables and therefore have to be encoded as factors. As an easy check, we can take the original corvariates data and check for the number of unique values in each raster. If the variable is continuous, we expect that there are a lot of different values - at maximum 1052 different values because we have that many entries. So, let’s have a look and assume that variables with 10 or less different values are categorical variables.

Code
cat_vars <- 
  covar_data |> 
  # Get number of distinct values per variable
  dplyr::summarise(dplyr::across(dplyr::everything(), ~ dplyr::n_distinct(.))) |> 
  # Turn df into long format for easy filtering
  tidyr::pivot_longer(dplyr::everything(), 
                      names_to = "variable", 
                      values_to = "n") |> 
  # Filter out variables with 10 or less distinct values
  dplyr::filter(n <= 10) |>
  # Extract the names of these variables
  dplyr::pull('variable')

cat("Variables with less than 10 distinct values:", 
    ifelse(length(cat_vars) == 0, "none", cat_vars))
Variables with less than 10 distinct values: none

Now that we have the names of the categorical values, we can mutate these columns in our df using the base function as.factor():

Code
final_data <- 
  final_data |> 
  dplyr::mutate(dplyr::across(cat_vars, ~ as.factor(.)))

5 Checking missing data

We are almost done with our data preparation, we just need to reduce it to sampling locations for which we have a decent amount of data on the covariates. Else, we blow up the model calibration with data that is not informative enough.

Code
# Get number of rows to calculate percentages
n_rows <- nrow(final_data)

# Get number of distinct values per variable
final_data |> 
  dplyr::summarise(dplyr::across(dplyr::everything(), 
                                 ~ length(.) - sum(is.na(.)))) |> 
  tidyr::pivot_longer(dplyr::everything(), 
                      names_to = "variable", 
                      values_to = "n") |>
  dplyr::mutate(perc_available = round(n / n_rows * 100)) |> 
  dplyr::arrange(perc_available) |> 
  head(10) |> 
  knitr::kable()
variable n perc_available
ph.30.50 856 81
ph.10.30 866 82
ph.50.100 859 82
timeset 871 83
ph.0.10 870 83
dclass 1006 96
site_id_unique 1052 100
x 1052 100
y 1052 100
dataset 1052 100

This looks good, we have no variable with a substantial amount of missing data. Generally, only pH measurements are lacking, which we should keep in mind when making predictions and inferences. Another great way to explore your data, is using the {visdat} package:

Code
final_data |> visdat::vis_miss()

Alright, we see that we are not missing any data in the covariate data. Mostly sampled data, specifically pH and timeset data is missing. We also see that this missing data is mostly from the same entry, so if we keep only entries where we have pH data - which is what we are interested here - we have a dataset with pracitally no missing data.

6 Save data

Code
saveRDS(final_data, 
        here::here("data/bern_sampling_locations_with_covariates.rds"))